and replace it with the internal, re-exported WindowHandleAccess
one.
Strictly speaking, this is a breaking change. In practice the
returned type of this trait was in `sixtyfps::re_exports`, so any
public use is questionable :)
That's all it is nowadays, it's a wrapper around Rc<Window>. It's not an
alias because we need to also "wrap" it to C++ via cbindgen, but that's
about it.
On Windows the default stack size is less than on Linux/macOS and
therefore an issues surfaces that is however operating system agnostic:
Due to inlining we generate quite big inner components with many fields.
Their field offsets are accumulated in the separate FIELD_OFFSETS
struct. The access pattern typically looks like this:
Self::FIELD_OFFSETS.some_field.apply_pin(...)
This is destructed into at least two smaller operations, fetching
some_field and then calling apply_pin with that field as self parameter.
In Rust debug builds the first operation, unfortunately, ends up making
a full copy of Self::FIELD_OFFSETS into a temporary variable on the
stack. Unfortunately also that stack space isn't reused, so every access
results in a new stack allocation to hold all of the FIELD_OFFSETS
structure (~5kb in the printer demo).
With so many accesses to that in the generated code, we run out of stack
space already in the new() function, in item_tree() as well as in the
Drop impl that also accesses all field offsets.
This patch applies the following workaround that does the trick. The
above expression is replaced with
{
let field = &Self::FIELD_OFFSETS.some_field;
*field
}.apply_pin(...)
Fixes#133
And do some passes before inlining
We will need the list of components before inlining in order to generate
them if we disable inlining
So we can do some passes on each component before they are inlining
I tried to put the flickable pass in that list, but it did not work
if the Flickable itself is the root of a component
The generated code tries to compute the box layout info for an empty
repeater, where we end up calling `&*std::begin(cells)` and MSVC
asserts that this is dereferencing an invalid iterator. As Olivier spotted,
`std::data` comes to the assert-free rescue :-)
Move those two classes into the private_api namespace, which is excluded
from the API reference documentation.
For generate code the explicit qualification of Property<T> is changed,
for the cbindgen generated item types the private_api namespace is
pulled into the cbindgen_private namespace.
this implies that we need to make sure the property are initialized in
order so that constant properties that depends on other constant properties
are correctly computed
